Handmade Code

We track identifiable AI-generated commits on public GitHub repositories — currently Claude Code and Codex, which leave traceable signatures. But these represent only the tip of the iceberg. Most AI-generated code is invisible: written through Copilot, Cursor, ChatGPT, and dozens of other tools that commit under the developer's name.

Traceable AI Code

—

identifiable on GitHub

Estimated Total AI

—

including hidden AI code

📈 AI Growth

—

traceable AI, 6 months

✋ Still Handmade

—

estimated human-only code

The Iceberg: What We Measure vs. What Exists

Our traceable data captures only tools that leave identifiable commit signatures — roughly 4% of all commits. But industry data tells a different story:

GitHub Copilot writes ~46% of its users' code (AI Expert Magazine, 2026)
Microsoft reports AI writes ~30% of their code (Tom's Hardware)
Industry surveys estimate 25–42% of all new code is AI-assisted (ShiftMag, Elite Brains)

Our conservative estimate: ~25% of all new code on GitHub is AI-generated — roughly 6× more than what's traceable.

~75% Human

~21% Hidden AI

~4%

Human-written AI-generated (untracked) AI-generated (traceable)

Methodology

We survey GitHub commit archives weekly, identifying AI-generated commits through traceable signatures:

Claude Code — identified via author-email: noreply@anthropic.com
Codex — identified via author-name: codex (community convention — likely undercounted)
Human — everything without an identifiable AI signature
Data from GitHub Search API, collected weekly since Aug 2025

Important caveat: Most AI coding tools (Copilot, Cursor, Windsurf, Aider, ChatGPT copy-paste) commit under the developer's name and are invisible to this method. Our traceable numbers represent a lower bound on AI-generated code.

The Iceberg: What We Measure vs. What Exists

Traceable AI Code Over Time

Weekly AI Commit Volume

Methodology